Skip to main content

Examples

Complete, runnable workflows for real jobs, in Python and TypeScript. Each page is self-contained and uses the real SDK surface end to end (Pareta.from_env() / Pareta.fromEnv()): call model="auto" with OpenAI-compatible inference, benchmark it on your data, and read the metered cost in dollars off run.cost. Grouped by what you are trying to do.

Use-case recipes​

Straight-up inference, one page per workload. Every recipe is the same shape — send your data to the interface that matches it (model="auto" for messages, dedicated routes for ranked lists, vectors, and audio) — and each section links to a full runnable program in Pareta-AI/examples.

  • Medical coding (ICD-10) — a clinical discharge summary in, ICD-10-CM codes as strict JSON out.
  • Retrieval: reranking and embeddings — the RAG stack: calibrated reranking, embedding-based semantic search, and the two-stage recall→precision pipeline.
  • Extraction: documents and contracts — pull JSON fields out of invoice images and PDFs (how to attach them to the prompt) and out of contract text.
  • Text classification — intent classification with a closed label set + few-shot, and hate-speech content moderation.
  • Summarization — meeting notes to an executive summary and action items, with a streaming variant.
  • Text to speech — synthesize spoken audio and save it to a file.
  • Speech to text — transcribe an audio clip from a path, raw bytes, or base64.

Prove it on your own data​

You have a job in plain English, or your own data, and want proof that "auto" wins on it.

  • Benchmark models on your own data — build an eval set from your rows, run "auto" against frontier="benchmarked" baselines, and read ranked results plus run.cost.
  • Document extraction (PDF/image) — the blob-task loop: build an eval set from your PDFs/images, upload_document per row, benchmark "auto" against vision frontier baselines, then run OpenAI-compatible inference with model="auto".

Inference patterns​

Getting tokens out efficiently.

  • Streaming chat completions — stream tokens with chat.completions.create(stream=True): iterate ChatCompletionChunk objects, read delta.content, accumulate full text, plus async streaming and metering behavior.
  • Concurrent calls with AsyncPareta — fire many inference and eval calls concurrently with AsyncPareta and asyncio.gather, with semaphore backpressure and per-task error handling.

Operate and monitor​

Watching what auto is doing, and what it costs.

  • Cost & quality monitoring — read what calls and eval runs cost, the projected savings vs frontier, and watch requests, success rate, and spend via auto.metrics().

Migrating in​

Already on the OpenAI SDK.

  • Migrating from the OpenAI SDK — keep using the openai client against Pareta (base_url + pareta_sk_ key), and when to switch to the pareta SDK for evals and auto's metrics.

See also​